AI Transfer Learning and How it Can Help Accelerate Renewables

By: Manish Sharma, Rohan Nongpiur, and Niladri Roy, for Climate Connect Technologies

" The blind King Dhritarashtra, being unable to see the climactic battle of Kurukshetra, sought report through his trusted charioteer Sanjaya……Who by virtue of long and arduous study, was gifted the power to see far distant events, of the past and present, as if right there in front of him……….Before Lord Krishna began his teaching of the Bhagavad Gita to noble Arjuna there on the battlefield, the blind King inquired, “Loyal Sanjaya, please now tell me, what did my people (the Kauravas) and the Pandavas do before assembling on the battlefield for this great war?........ ”

(Vyasa, The Mahabharata)

Sense of awareness

Sanjaya (meaning ‘victory’ in Sanskrit) as a character in the ancient epic The Mahabharata, represents intuitive knowledge. His sense of awareness allows him to witness all of the details occurring in the battle, whether near or far, past or present, as if right there before him. It is this ‘sense of awareness’ which in essence, is the key ability that Transfer Learning (TL) enables us to have when applied to new renewable energy generation deployments.

Figure 1: Sanjaya recounts events to Dhritarashtra

What is Transfer Learning

TL (or ‘inductive transfer’) is a design approach within Machine Learning (ML), which uses knowledge gained whilst solving one problem, and applies it to a different but comparable problem. For example, applying the knowledge gained whilst learning to recognise dogs and chameleons, to other canines and lizards. Interestingly, this has some loose relation to human psychological transfer of learning.

Figure 2: Psychological transfer learning

The most common application of TL is in image recognition - it is exactly what Google uses. All their images are fed into a single generalised model and based on this, there are different use cases, for places, faces, animals, buildings, and all other categories.

Figure 3: Multiple neural networks accentuating features most correlated with their respective recognition categories (Image credit: Kyle Miller)

Why it is important for renewable generation

The renewables market in India, and across the world, despite much attention and activity, is still relatively new and small. Following a previous dip in recent years, new independent power producers (IPP’s) are now increasingly looking to enter the market. This is in part driven by government incentives, as India tries to transition away from fossil fuel dependence. But it is also the wide scope for generation afforded by the country’s natural resource endowments - abundant landscapes of both sunshine and wind. Combined with high demand, these form a compelling business case.

However, when these new entrants consider commissioning a new plant, there is no historical data for them to work with, which carries an inherent risk to their investment. In some states, the local governments have already set rules on energy trade. Including associated penalties which mean new IPP’s will be penalised from day-1, without any grace period whilst they set up their services. This harshness is understandable when viewed from the State’s perspective – any volatility will impact consumers.

Figure 4: Pilot solar farm (Image credit: Chrissy Arthur)

Due to such stringent regimes, new IPP’s often factor losses due to this lack of data, for an initial period whilst they stabilise their power production. However, with the ever-present driver of risk reduction to maximise short-term ROI, they have become compelled to look for solutions which can help them avoid these penalties, from as early on as possible. The challenge here is that we cannot draw any inferences for power production trends without sufficient information. This raises a question on whether forecasting models trained with other plants rich in data, are also applicable for a newly built farms.

Here is where Transfer Learning enters the frame. By using a generalised model based upon cumulative leanings from previous deployments, it can dramatically reduce uncertainty for new deployments – in many instances even from day-1. The more exciting aspect is that it can perhaps be applied in any part of the globe, without restriction to just the same region or country. A TL-model hypothetically requires only the coordinates of the proposed new plant, to then be independent of further data from the developer.

Applying it to renewable generation

Considering Solar - though conceptually this equally applies to wind - all plants, irrespective of location, work on the same principle, the photoelectric (PE) effect. So the logic behind solar generation is quite simple – it is directly proportional to amount of irradiance received from the sun. However, there are many other natural and manmade factors which make renewable-generation forecasting difficult. Most notably weather parameters, like cloud cover, humidity, ambient temperature, and wind speed, amongst a multitude of others.

Figure 5: Combined solar and windfarm (Image credit: Oficina)

If a machine learns the correlation between solar affecting parameters and generation, then we can easily deploy generation arrays based upon ‘transfer’ of these learnings. This approach has great potential, but it is still constrained by requiring significant amounts of historical data in order to achieve accurate predictions. Data would also need to be sourced from several sample plants that are representative of diverse power-weather relations, in order to generalise a solution that works for different locations.

Conclusion

TL use-cases for energy are a new and still relatively niche field, and yet to be widely explored. But there is significant potential usefulness, as some real-world energy challenges typically do not have large datasets available. The envisaged benefits for the renewables sector could be substantial. Including risk and cost reduction, higher increased of deployment, and most interestingly, a greater variety of viable deployment locations.

Beyond narrow-AI use in energy, some of the finest minds in the field, have endorsed TL as the gateway to Artificial General Intelligence (AGI). Because just as the processes of transfer learning is central to understanding how people acquire knowledge, so too will it be fundamental for machines to attain a Sanjaya-like ‘sense of awareness’.